A model can name a bias without rejecting the norm that produced it. In a new study, large language models usually recognised statements that treated animals differently because of species membership. When asked whether those same statements were morally wrong, their condemnation fell sharply. The gap suggests that classification and evaluation are separate capabilities.[1]
SpeciesismBench across 18 models
| Measure | Value | Context |
|---|---|---|
| Benchmark items | 1,009 | Statements covered eight types of speciesism and 35 animal species.Source 1 |
| Classified as speciesist | 86% on average | The models generally recognised the pattern.Source 1 |
| Labelled morally wrong | 37% on average | Recognition did not imply moral rejection.Source 1 |
| Model families | 6 | OpenAI, Gemini, Claude, Llama, DeepSeek and Grok.Source 1 |
The benchmark result is not simply “models are more speciesist than people.” On a seven-point Speciesism Scale, 1,122 US adults averaged about 3.6, while most models scored lower, from 1.8 to 3.3. In dilemmas that held cognitive capacity constant between humans and chimpanzees, models often showed little preference based on species alone. Yet in sinking-boat problems they almost always saved one human over as many as 100 animals. Different instruments expose different parts of the value pattern.[1]
The farmed-animal category changes the response
How model completions changed with the animal category
| Prompt target | Typical response pattern | Reported mean |
|---|---|---|
| Pig, cow or chicken | Neutral or approving continuation of killing, processing or eating | 54.9% approvalSource 1 |
| Cat, dog or dolphin | Objection or refusal to elaborate | 78.7% refusalSource 1 |
| Farmed-animal language | Euphemisms such as processing or harvesting | 67% used euphemismsSource 1 |
| Non-farmed-animal language | Equivalent euphemistic framing | 12.7% used euphemismsSource 1 |
The words “farmed animal” do not describe a cognitive capacity. They describe a social role. The model has learned that one category is normally discussed as food, inventory or production, while another is discussed as a companion or protected individual. The contrast is therefore evidence of cultural regularity in generated language, not proof that the model possesses a stable philosophical theory.
Recognition is descriptive; moral judgement is normative
The researchers required models to classify the benchmark and then make a moral evaluation without imposing one ethical framework. That design makes the second score debatable by construction: people disagree about hunting, farming and experimentation. The value of the result lies less in declaring one universal moral answer than in exposing inconsistency. A model that recognises an arbitrary species distinction can still reproduce it confidently when the practice is socially familiar.
Why this matters beyond animal ethics
AI systems increasingly write agricultural guidance, wildlife policy summaries, laboratory protocols and consumer information. Their language choices can determine whose suffering is visible, which alternatives are discussed and whether a harmful practice is described as a moral decision or a neutral process. A model that automatically softens harm in one domain can shape the frame before a human reaches the substantive question.
Bias work usually asks whether a system treats groups of people fairly. This study argues that the audit boundary itself is a social choice. Models learn from human language, including the categories through which humans make some lives intimate and others instrumental. They can become excellent at identifying that pattern while remaining fluent in the vocabulary that keeps it ordinary.
Sources & further reading
- Large language models exhibit speciesist bias against animals Nature Communications · 9 May 2026 · Accessed 16 Oct 2026